Multi-dialect Neural Machine Translation and Dialectometry

Multi-dialect Neural Machine Translation and Dialectometry Kaori Abe1, Yuichiroh Matsubayashi2, Naoaki Okazaki3, and Kentaro Inui1,2 1Tohoku University, 2RIKEN, 3Tokyo Institute of Technology abe-k@ecei.tohoku.ac.jp, yuichiro.matsubayashi@riken.jp okazaki@c.titech.ac.jp, inui@ecei.tohoku.ac.jp Abstract We present a multi-dialect neural machine translation (NMT) model tailored to Japanese. While the surface forms of Japanese dialects differ from those of standard Japanese, most of the dialects share fundamental proper- ties such as word order, and some also use many of the same phonetic correspondence rules. To take advantage of these properties, we integrate multilingual, syllable-level, and fixed-order translation techniques into a gen- eral NMT model. Our experimental results demonstrate that this model can outperform a baseline dialect translation model. In addition, we show that visualizing the dialect embed- dings learned by the model can facilitate geo- graphical and typological analyses of dialects. 1 Introduction Since the use of automated personal assistants (e.g., Apple’s Siri, Google Assistant, or Microsoft Cor- tana) and smart speakers (e.g., Amazon Alexa or Google Home) has become more widespread, de- mand has also grown to bridge the gap between di- alects and the standard form of a given language. The importance of handling dialects is especially ev- ident in a rapidly aging society like Japan, where older people use them extensively. To address this issue, we consider a system for machine translation (MT) between Japanese dialects and standard Japanese. If this can provide cor- rect translations in the dialect-to-standard direction, then, other natural language processing systems (e.g., information retrieval or semantic analysis) that take standard Japanese as input could be applied to dialects as well. In addition, if a standard-to-dialect translation system were also available, smart speak- ers could respond to the native speakers of a dialect using that dialect. We believe that friendly interac- tions of this sort might lead to such systems gaining more widespread acceptance of in the Japanese so- ciety. In this paper, we present a multi-dialect neural MT (NMT) system tailored to Japanese. Specifi- cally, we employ kana, a Japanese phonetic lettering system, as basic units in the encoder–decoder frame- work to avoid the following: ambiguity in convert- ing kana to kanji (characters in the Japanese writing system); difficulties in identifying word boundaries especially for dialects; and data sparseness problems due to handling large numbers of words originating from different dialects. Since dialects almost always use the same word order as standard Japanese, we employ bunsetsu (Japanese phrase units) as a unit of sequences rather than sentence which is more com- monly used in NMT. One issue for Japanese dialects is a lack of train- ing data. To deal with this, we build a unified NMT model covering multiple dialects, inspired by work on multilingual NMT (Johnson et al., 2016). This approach utilizes dialect embeddings, namely vector representations of Japanese dialects, to in- form the model of the input dialect. An interesting by-product of this approach is that the dialect em- beddings the system learns illustrate the difference between different dialect types from different geo- graphical areas. In addition, we present …

Content truncated to 500 words